
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_5)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(tmp) end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(tmp); end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
Herbie found 5 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_5)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(tmp) end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(tmp); end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (fabs (* (floor h) (* -1.0 (* dX.u (* dY.v (floor w)))))))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor w) dY.u))
(t_3 (* t_2 t_2))
(t_4 (* dY.u (floor w)))
(t_5 (* (floor w) dX.u))
(t_6 (* (floor h) dX.v))
(t_7 (* dX.u (floor w)))
(t_8 (* (* t_7 dX.u) (floor w)))
(t_9 (fabs (* (floor h) (- (* t_4 dX.v) (* dY.v t_7)))))
(t_10 (* (floor h) dY.v))
(t_11 (fmax (+ (* t_5 t_5) (* t_6 t_6)) (+ t_3 (* t_10 t_10))))
(t_12 (fabs (- (* t_5 t_10) (* t_6 t_2))))
(t_13 (sqrt t_11))
(t_14
(fmax
(fma (* (* dY.v dY.v) (floor h)) (floor h) t_3)
(fma (* t_1 dX.v) (floor h) t_8)))
(t_15 (sqrt t_14))
(t_16 (* t_1 t_1))
(t_17
(fmax
(fma
(* t_4 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))
(/ (- (* t_8 t_8) (* t_16 t_16)) (- t_8 t_16))))
(t_18 (sqrt t_17)))
(if (<=
(log2
(if (> (/ t_11 t_12) (floor maxAniso))
(/ t_13 (floor maxAniso))
(/ t_12 t_13)))
100.0)
(log2
(if (> (/ t_14 t_9) (floor maxAniso))
(/ t_15 (floor maxAniso))
(/ t_9 t_15)))
(log2
(if (> (/ t_17 t_0) (floor maxAniso))
(/ t_18 (floor maxAniso))
(/ t_0 t_18))))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = fabsf((floorf(h) * (-1.0f * (dX_46_u * (dY_46_v * floorf(w))))));
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(w) * dY_46_u;
float t_3 = t_2 * t_2;
float t_4 = dY_46_u * floorf(w);
float t_5 = floorf(w) * dX_46_u;
float t_6 = floorf(h) * dX_46_v;
float t_7 = dX_46_u * floorf(w);
float t_8 = (t_7 * dX_46_u) * floorf(w);
float t_9 = fabsf((floorf(h) * ((t_4 * dX_46_v) - (dY_46_v * t_7))));
float t_10 = floorf(h) * dY_46_v;
float t_11 = fmaxf(((t_5 * t_5) + (t_6 * t_6)), (t_3 + (t_10 * t_10)));
float t_12 = fabsf(((t_5 * t_10) - (t_6 * t_2)));
float t_13 = sqrtf(t_11);
float t_14 = fmaxf(fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), t_3), fmaf((t_1 * dX_46_v), floorf(h), t_8));
float t_15 = sqrtf(t_14);
float t_16 = t_1 * t_1;
float t_17 = fmaxf(fmaf((t_4 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))), (((t_8 * t_8) - (t_16 * t_16)) / (t_8 - t_16)));
float t_18 = sqrtf(t_17);
float tmp;
if ((t_11 / t_12) > floorf(maxAniso)) {
tmp = t_13 / floorf(maxAniso);
} else {
tmp = t_12 / t_13;
}
float tmp_2;
if (log2f(tmp) <= 100.0f) {
float tmp_3;
if ((t_14 / t_9) > floorf(maxAniso)) {
tmp_3 = t_15 / floorf(maxAniso);
} else {
tmp_3 = t_9 / t_15;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_17 / t_0) > floorf(maxAniso)) {
tmp_4 = t_18 / floorf(maxAniso);
} else {
tmp_4 = t_0 / t_18;
}
tmp_2 = log2f(tmp_4);
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = abs(Float32(floor(h) * Float32(Float32(-1.0) * Float32(dX_46_u * Float32(dY_46_v * floor(w)))))) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(t_2 * t_2) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(dX_46_u * floor(w)) t_8 = Float32(Float32(t_7 * dX_46_u) * floor(w)) t_9 = abs(Float32(floor(h) * Float32(Float32(t_4 * dX_46_v) - Float32(dY_46_v * t_7)))) t_10 = Float32(floor(h) * dY_46_v) t_11 = fmax(Float32(Float32(t_5 * t_5) + Float32(t_6 * t_6)), Float32(t_3 + Float32(t_10 * t_10))) t_12 = abs(Float32(Float32(t_5 * t_10) - Float32(t_6 * t_2))) t_13 = sqrt(t_11) t_14 = fmax(fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), t_3), fma(Float32(t_1 * dX_46_v), floor(h), t_8)) t_15 = sqrt(t_14) t_16 = Float32(t_1 * t_1) t_17 = fmax(fma(Float32(t_4 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))), Float32(Float32(Float32(t_8 * t_8) - Float32(t_16 * t_16)) / Float32(t_8 - t_16))) t_18 = sqrt(t_17) tmp = Float32(0.0) if (Float32(t_11 / t_12) > floor(maxAniso)) tmp = Float32(t_13 / floor(maxAniso)); else tmp = Float32(t_12 / t_13); end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(100.0)) tmp_3 = Float32(0.0) if (Float32(t_14 / t_9) > floor(maxAniso)) tmp_3 = Float32(t_15 / floor(maxAniso)); else tmp_3 = Float32(t_9 / t_15); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_17 / t_0) > floor(maxAniso)) tmp_4 = Float32(t_18 / floor(maxAniso)); else tmp_4 = Float32(t_0 / t_18); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
t_0 := \left|\left\lfloor h\right\rfloor \cdot \left(-1 \cdot \left(dX.u \cdot \left(dY.v \cdot \left\lfloor w\right\rfloor \right)\right)\right)\right|\\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := t\_2 \cdot t\_2\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_8 := \left(t\_7 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_9 := \left|\left\lfloor h\right\rfloor \cdot \left(t\_4 \cdot dX.v - dY.v \cdot t\_7\right)\right|\\
t_10 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_11 := \mathsf{max}\left(t\_5 \cdot t\_5 + t\_6 \cdot t\_6, t\_3 + t\_10 \cdot t\_10\right)\\
t_12 := \left|t\_5 \cdot t\_10 - t\_6 \cdot t\_2\right|\\
t_13 := \sqrt{t\_11}\\
t_14 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , t\_3\right), \mathsf{fma}\left(t\_1 \cdot dX.v, \left\lfloor h\right\rfloor , t\_8\right)\right)\\
t_15 := \sqrt{t\_14}\\
t_16 := t\_1 \cdot t\_1\\
t_17 := \mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right), \frac{t\_8 \cdot t\_8 - t\_16 \cdot t\_16}{t\_8 - t\_16}\right)\\
t_18 := \sqrt{t\_17}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_11}{t\_12} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_13}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_12}{t\_13}\\
\end{array} \leq 100:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_14}{t\_9} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_15}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_9}{t\_15}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_17}{t\_0} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_18}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_18}\\
\end{array}\\
\end{array}
if (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) < 100Initial program 75.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
if 100 < (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) Initial program 75.8%
Applied rewrites75.8%
Taylor expanded in dX.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f3275.0%
Applied rewrites75.0%
Taylor expanded in dX.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f3274.9%
Applied rewrites74.9%
lift-fma.f32N/A
+-commutativeN/A
flip-+N/A
lower-unsound-/.f32N/A
Applied rewrites71.6%
lift-fma.f32N/A
+-commutativeN/A
flip-+N/A
lower-unsound-/.f32N/A
Applied rewrites45.6%
lift-fma.f32N/A
+-commutativeN/A
flip-+N/A
lower-unsound-/.f32N/A
Applied rewrites48.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* t_0 t_0))
(t_2 (* dY.u (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5
(fmax
(fma
(* t_2 dY.u)
(floor w)
(exp (* (log (* (- dY.v) (floor h))) 2.0)))
(* (pow dX.v 2.0) (pow (floor h) 2.0))))
(t_6 (sqrt t_5))
(t_7 (* dX.u (floor w)))
(t_8 (fabs (* (floor h) (- (* t_2 dX.v) (* dY.v t_7)))))
(t_9 (* (floor h) dY.v))
(t_10 (fmax (+ (* t_3 t_3) (* t_4 t_4)) (+ t_1 (* t_9 t_9))))
(t_11 (fabs (- (* t_3 t_9) (* t_4 t_0))))
(t_12 (sqrt t_10))
(t_13
(fmax
(fma (* (* dY.v dY.v) (floor h)) (floor h) t_1)
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_7 dX.u) (floor w)))))
(t_14 (sqrt t_13)))
(if (<=
(log2
(if (> (/ t_10 t_11) (floor maxAniso))
(/ t_12 (floor maxAniso))
(/ t_11 t_12)))
100.0)
(log2
(if (> (/ t_13 t_8) (floor maxAniso))
(/ t_14 (floor maxAniso))
(/ t_8 t_14)))
(log2
(if (> (/ t_5 t_8) (floor maxAniso))
(/ t_6 (floor maxAniso))
(/ t_8 t_6))))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = t_0 * t_0;
float t_2 = dY_46_u * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = fmaxf(fmaf((t_2 * dY_46_u), floorf(w), expf((logf((-dY_46_v * floorf(h))) * 2.0f))), (powf(dX_46_v, 2.0f) * powf(floorf(h), 2.0f)));
float t_6 = sqrtf(t_5);
float t_7 = dX_46_u * floorf(w);
float t_8 = fabsf((floorf(h) * ((t_2 * dX_46_v) - (dY_46_v * t_7))));
float t_9 = floorf(h) * dY_46_v;
float t_10 = fmaxf(((t_3 * t_3) + (t_4 * t_4)), (t_1 + (t_9 * t_9)));
float t_11 = fabsf(((t_3 * t_9) - (t_4 * t_0)));
float t_12 = sqrtf(t_10);
float t_13 = fmaxf(fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), t_1), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_7 * dX_46_u) * floorf(w))));
float t_14 = sqrtf(t_13);
float tmp;
if ((t_10 / t_11) > floorf(maxAniso)) {
tmp = t_12 / floorf(maxAniso);
} else {
tmp = t_11 / t_12;
}
float tmp_2;
if (log2f(tmp) <= 100.0f) {
float tmp_3;
if ((t_13 / t_8) > floorf(maxAniso)) {
tmp_3 = t_14 / floorf(maxAniso);
} else {
tmp_3 = t_8 / t_14;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_5 / t_8) > floorf(maxAniso)) {
tmp_4 = t_6 / floorf(maxAniso);
} else {
tmp_4 = t_8 / t_6;
}
tmp_2 = log2f(tmp_4);
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(t_0 * t_0) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = fmax(fma(Float32(t_2 * dY_46_u), floor(w), exp(Float32(log(Float32(Float32(-dY_46_v) * floor(h))) * Float32(2.0)))), Float32((dX_46_v ^ Float32(2.0)) * (floor(h) ^ Float32(2.0)))) t_6 = sqrt(t_5) t_7 = Float32(dX_46_u * floor(w)) t_8 = abs(Float32(floor(h) * Float32(Float32(t_2 * dX_46_v) - Float32(dY_46_v * t_7)))) t_9 = Float32(floor(h) * dY_46_v) t_10 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4)), Float32(t_1 + Float32(t_9 * t_9))) t_11 = abs(Float32(Float32(t_3 * t_9) - Float32(t_4 * t_0))) t_12 = sqrt(t_10) t_13 = fmax(fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), t_1), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_7 * dX_46_u) * floor(w)))) t_14 = sqrt(t_13) tmp = Float32(0.0) if (Float32(t_10 / t_11) > floor(maxAniso)) tmp = Float32(t_12 / floor(maxAniso)); else tmp = Float32(t_11 / t_12); end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(100.0)) tmp_3 = Float32(0.0) if (Float32(t_13 / t_8) > floor(maxAniso)) tmp_3 = Float32(t_14 / floor(maxAniso)); else tmp_3 = Float32(t_8 / t_14); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_5 / t_8) > floor(maxAniso)) tmp_4 = Float32(t_6 / floor(maxAniso)); else tmp_4 = Float32(t_8 / t_6); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := t\_0 \cdot t\_0\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dY.u, \left\lfloor w\right\rfloor , e^{\log \left(\left(-dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot 2}\right), {dX.v}^{2} \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right)\\
t_6 := \sqrt{t\_5}\\
t_7 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_8 := \left|\left\lfloor h\right\rfloor \cdot \left(t\_2 \cdot dX.v - dY.v \cdot t\_7\right)\right|\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_4 \cdot t\_4, t\_1 + t\_9 \cdot t\_9\right)\\
t_11 := \left|t\_3 \cdot t\_9 - t\_4 \cdot t\_0\right|\\
t_12 := \sqrt{t\_10}\\
t_13 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , t\_1\right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_7 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_14 := \sqrt{t\_13}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_10}{t\_11} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_12}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_11}{t\_12}\\
\end{array} \leq 100:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_13}{t\_8} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_14}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{t\_14}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_5}{t\_8} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_6}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{t\_6}\\
\end{array}\\
\end{array}
if (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) < 100Initial program 75.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
if 100 < (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) Initial program 75.8%
Applied rewrites75.8%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3263.9%
Applied rewrites63.9%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3261.8%
Applied rewrites61.8%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-unsound-exp.f32N/A
lower-unsound-*.f32N/A
lower-unsound-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lower-*.f32N/A
lower-neg.f3265.5%
Applied rewrites65.5%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3254.2%
Applied rewrites54.2%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3250.4%
Applied rewrites50.4%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3252.8%
Applied rewrites52.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* dX.u (floor w)))
(t_2
(fmax
(fma (* (* dY.v dY.v) (floor h)) (floor h) (* t_0 t_0))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_1 dX.u) (floor w)))))
(t_3 (sqrt t_2))
(t_4
(fabs (* (floor h) (- (* (* dY.u (floor w)) dX.v) (* dY.v t_1))))))
(log2
(if (> (/ t_2 t_4) (floor maxAniso))
(/ t_3 (floor maxAniso))
(/ t_4 t_3)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = dX_46_u * floorf(w);
float t_2 = fmaxf(fmaf(((dY_46_v * dY_46_v) * floorf(h)), floorf(h), (t_0 * t_0)), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_1 * dX_46_u) * floorf(w))));
float t_3 = sqrtf(t_2);
float t_4 = fabsf((floorf(h) * (((dY_46_u * floorf(w)) * dX_46_v) - (dY_46_v * t_1))));
float tmp;
if ((t_2 / t_4) > floorf(maxAniso)) {
tmp = t_3 / floorf(maxAniso);
} else {
tmp = t_4 / t_3;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(dX_46_u * floor(w)) t_2 = fmax(fma(Float32(Float32(dY_46_v * dY_46_v) * floor(h)), floor(h), Float32(t_0 * t_0)), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_1 * dX_46_u) * floor(w)))) t_3 = sqrt(t_2) t_4 = abs(Float32(floor(h) * Float32(Float32(Float32(dY_46_u * floor(w)) * dX_46_v) - Float32(dY_46_v * t_1)))) tmp = Float32(0.0) if (Float32(t_2 / t_4) > floor(maxAniso)) tmp = Float32(t_3 / floor(maxAniso)); else tmp = Float32(t_4 / t_3); end return log2(tmp) end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , t\_0 \cdot t\_0\right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left\lfloor h\right\rfloor \cdot \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.v - dY.v \cdot t\_1\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_4} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_3}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_3}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lower-fma.f3275.8%
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (fabs (* (floor h) (- (* t_0 dX.v) (* dY.v t_1)))))
(t_3
(fmax
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_1 dX.u) (floor w)))))
(t_4 (sqrt t_3)))
(log2
(if (> (/ t_3 t_2) (floor maxAniso))
(/ t_4 (floor maxAniso))
(/ t_2 t_4)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = fabsf((floorf(h) * ((t_0 * dX_46_v) - (dY_46_v * t_1))));
float t_3 = fmaxf(fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_1 * dX_46_u) * floorf(w))));
float t_4 = sqrtf(t_3);
float tmp;
if ((t_3 / t_2) > floorf(maxAniso)) {
tmp = t_4 / floorf(maxAniso);
} else {
tmp = t_2 / t_4;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = abs(Float32(floor(h) * Float32(Float32(t_0 * dX_46_v) - Float32(dY_46_v * t_1)))) t_3 = fmax(fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_1 * dX_46_u) * floor(w)))) t_4 = sqrt(t_3) tmp = Float32(0.0) if (Float32(t_3 / t_2) > floor(maxAniso)) tmp = Float32(t_4 / floor(maxAniso)); else tmp = Float32(t_2 / t_4); end return log2(tmp) end
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left|\left\lfloor h\right\rfloor \cdot \left(t\_0 \cdot dX.v - dY.v \cdot t\_1\right)\right|\\
t_3 := \mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_4 := \sqrt{t\_3}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_3}{t\_2} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_4}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1
(fmax
(fma t_0 t_0 (* (* (* dY.u (floor w)) dY.u) (floor w)))
(fma
(* (floor h) (floor h))
(* dX.v dX.v)
(* (* (* dX.u (floor w)) dX.u) (floor w)))))
(t_2 (sqrt t_1))
(t_3 (fabs (* (* (* (floor h) dY.u) (floor w)) dX.v))))
(log2
(if (> (/ t_1 t_3) (floor maxAniso))
(/ t_2 (floor maxAniso))
(/ t_3 t_2)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_v * floorf(h);
float t_1 = fmaxf(fmaf(t_0, t_0, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))), fmaf((floorf(h) * floorf(h)), (dX_46_v * dX_46_v), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))));
float t_2 = sqrtf(t_1);
float t_3 = fabsf((((floorf(h) * dY_46_u) * floorf(w)) * dX_46_v));
float tmp;
if ((t_1 / t_3) > floorf(maxAniso)) {
tmp = t_2 / floorf(maxAniso);
} else {
tmp = t_3 / t_2;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = fmax(fma(t_0, t_0, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))), fma(Float32(floor(h) * floor(h)), Float32(dX_46_v * dX_46_v), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)))) t_2 = sqrt(t_1) t_3 = abs(Float32(Float32(Float32(floor(h) * dY_46_u) * floor(w)) * dX_46_v)) tmp = Float32(0.0) if (Float32(t_1 / t_3) > floor(maxAniso)) tmp = Float32(t_2 / floor(maxAniso)); else tmp = Float32(t_3 / t_2); end return log2(tmp) end
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , dX.v \cdot dX.v, \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_2 := \sqrt{t\_1}\\
t_3 := \left|\left(\left(\left\lfloor h\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot dX.v\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_1}{t\_3} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_2}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{t\_2}\\
\end{array}
\end{array}
Initial program 75.8%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3274.9%
Applied rewrites74.9%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3274.8%
Applied rewrites74.8%
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites74.8%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
swap-sqrN/A
lower-fma.f32N/A
lower-*.f32N/A
lower-*.f3274.8%
Applied rewrites74.8%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
swap-sqrN/A
lower-fma.f32N/A
lower-*.f32N/A
lower-*.f3274.8%
Applied rewrites74.8%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
swap-sqrN/A
lower-fma.f32N/A
lower-*.f32N/A
lower-*.f3274.8%
Applied rewrites74.8%
herbie shell --seed 2025187
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (LOD)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(log2 (if (> (/ (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u))))) (floor maxAniso)) (/ (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v))))) (floor maxAniso)) (/ (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u)))) (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))))))